Abstract: Mining potential associations among entities in ethnic cultural knowledge graphs was considered valuable for revealing evolutionary relationships among cultural elements and for providing new technical paths for systematic modeling and intelligent reasoning. However, because such graphs were characterized by many-to-one attribute affiliations, highly homogeneous structures, and numerous noisy edges, existing link prediction models were limited in simultaneously capturing fine-grained feature associations among cultural entities and resisting noise interference in the graphs, which restricted predictive performance and model robustness. To address these problems, a graph neural network model with similarity-aware features, SRGCN (Similarity-aware Relational Graph Convolutional Network) was proposed. SRGCN constructed a dynamic aggregation mechanism on the basis of node feature similarity to capture associations among entities more accurately; at the same time, a dual-level contrastive learning framework was introduced to suppress noise interference effectively, and a linearly weighted multi-objective loss function was designed, where the weights of primary and auxiliary tasks were dynamically adjusted to further enhance robustness. Experimental results on the HeritEdge ethnic cultural knowledge graph dataset show that SRGCN outperforms the strongest baseline model, LTRGN (Linear self-attention with multi-Relational Graph Network), achieving relative improvements of 28.4% and 32.5% on Mean Reciprocal Rank (MRR) and Hits@10, respectively, and delivers better overall performance.